Predictive trajectory generation methods, equipment, vehicles, and storage media

Abstract

This application provides a method, device, vehicle, and storage medium for predicting a trajectory. The method includes: predicting noise in the environmental perception information based on the vehicle's environmental perception information; removing the noise based on the environmental perception information to generate multiple sets of anchor points; the multiple sets of anchor points representing multiple predicted trajectories in multiple directions; and determining a final predicted trajectory from the multiple predicted trajectories. This achieves multimodal operation, allowing the output of a final predicted trajectory from multiple predicted trajectories.

Description

Technical Field

[0001] This invention relates to the field of vehicle control technology, and in particular to a method, device, vehicle, and storage medium for predictive trajectory generation. Background Technology

[0002] As driver assistance systems evolve from Level 2+ (Level 2 Partial Driving Automation) to Level 4 (Level 4 High Driving Automation), trajectory planning modules have gradually shifted from rule-driven to deep learning-based evolution.

[0003] In complex urban roads, highly dynamic environments, and scenarios with guide lines, the trajectory prediction methods of related technologies may fail during execution if the input samples of the trajectory generation model are noisy. This can cause the output trajectories to all converge to a single trajectory in the direction of the noisy guide line, or the trajectory prediction may exit due to the inability to output execution results. Summary of the Invention

[0004] This application provides an improved method, apparatus, vehicle, and storage medium for predictive trajectory generation.

[0005] This application provides a trajectory generation method, including: Based on the vehicle's environmental perception information, predict the noise level of the environmental perception information; Based on the environmental perception information, the noise is removed, and multiple sets of anchor points are generated; the multiple sets of anchor points are used to represent multiple predicted trajectories in multiple directions; The final predicted trajectory is determined from multiple predicted trajectories.

[0006] This application provides an electronic device including one or more processors for implementing the method described in any of the preceding claims.

[0007] This application provides a vehicle including one or more processors for implementing the method as described in any of the preceding claims.

[0008] This application provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the method described in any of the preceding claims.

[0009] This application provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the method described in any of the preceding claims.

[0010] In some embodiments, the predicted trajectory generation method of this application predicts noise in the environmental perception information based on the vehicle's environmental perception information, and removes the noise based on the environmental perception information to generate multiple sets of anchor points. These multiple sets of anchor points are used to represent multiple predicted trajectories in multiple directions, thereby achieving multimodal processing. Thus, based on the identified noise, predicted trajectories in multiple directions can be output normally, and the final predicted trajectory can be output from the multiple predicted trajectories. This improves the anti-interference capability and scene adaptability of the trajectory prediction process, ensuring the stability of assisted driving decisions. Attached Figure Description

[0011] Figure 1 The diagram shown is a schematic representation of the trajectory generation model used in the trajectory generation method provided in this embodiment of the application. Figure 2 The diagram shown is a flowchart illustrating the predicted trajectory generation method provided in an embodiment of this application. Figure 3 The diagram shown is a flowchart illustrating the predicted trajectory generation method provided in this application, which includes a dynamic sector feasible region. Figure 4 The diagram shown is a structural schematic of an electronic device provided in an embodiment of this application. Detailed Implementation

[0012] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0013] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0014] In the trajectory prediction methods of related technologies, if the input samples of the trajectory generation model are noisy during the execution process, the trajectory generation model cannot identify the noise, causing the output trajectories to all converge to a single trajectory in the direction of the noise-guided line, or it will exit trajectory prediction after failing to output the execution result.

[0015] To address the technical problem in related technologies where the trajectory generation model cannot identify noise, resulting in all output trajectories converging to a single trajectory in the direction of the noise-guided line, or exiting trajectory prediction due to the inability to output execution results, this application provides a predicted trajectory generation method. Based on the vehicle's environmental perception information, the method predicts the noise in the environmental perception information, and then removes the noise based on the environmental perception information, generating multiple sets of anchor points. These multiple sets of anchor points represent multiple predicted trajectories in multiple directions, thereby achieving multimodal prediction. Thus, based on noise identification, predicted trajectories in multiple directions can be output normally, and the final predicted trajectory can be output from the multiple predicted trajectories. This improves the anti-interference capability and scene adaptability of the trajectory prediction process, ensuring the stability of assisted driving decision-making.

[0016] Furthermore, the embodiments of this application eliminate the interference of noise on trajectory prediction by predicting, identifying, and removing noise in the environmental perception information in advance. Simultaneously, the multimodal trajectory generated based on multiple sets of anchor points can retain trajectory candidates in multiple directions, avoiding the output of a single, convergent, and unreasonable trajectory. In this way, an environmental perception-driven intention-based trajectory generation strategy can be implemented, independent of guide lines.

[0017] Next, the trajectory generation method based on guide lines in related technologies constructs a spatiotemporal graph by using guide lines as graph nodes and combining them with surrounding obstacles. It then uses a Transformer or GNN (Graph Neural Network) to predict multiple future trajectories. However, these related technologies rely solely on guide lines as the sole source of intent.

[0018] When the guide line fails, such as in a construction zone or when the lane lines are blocked, it will cause the mode to collapse, and the output trajectory will all converge to a single trajectory in the direction of the noise guide line, reducing the robustness of the system.

[0019] To address the technical problem of guide line failure and reduced system robustness in the aforementioned related technologies, the predicted trajectory generation method provided in this application generates multiple sets of anchor points by removing noise based on environmental perception information. These multiple sets of anchor points are used to represent multiple predicted trajectories in multiple directions, enabling environmental perception-driven intention trajectory generation without relying on guide line strategies or multimodal excitation mechanisms.

[0020] Figure 1 The diagram shown is a schematic representation of the trajectory generation model used in the trajectory generation method provided in this embodiment of the application.

[0021] like Figure 1As shown, the trajectory generation model applied by the predicted trajectory generation method provided in this application embodiment includes, but is not limited to, a Driver Style Embedding Extractor (DSEE), a Virtual Waypoint Generator with Style Modulation (VWG-SM), a Dynamic Fan-shaped Feasible Region (DFFR), and a Personalized Trajectory Selector with Semantic Labeling (PTS-SL).

[0022] Among them, the driving style embedding extraction module DSEE is used to extract driving style-related embedding features from the input environmental perception features E, providing a basis for personalized conditional modulation in the subsequent generation process.

[0023] The aiming point generation and modulation module VWG-SM generates multiple candidate estimation sequences based on the input environmental perception features and the extracted driving style embedding through multimodal conditional diffusion. During the generation process, driving style conditions are injected into the Transformer network through conditional layer normalization to ensure that the generated results match the target driving preferences.

[0024] The aforementioned dynamic sector spatiotemporal hard constraint module DFFR constructs a dynamic sector feasible region, selects candidate estimates that meet vehicle kinematic constraints and environmental obstacle constraints, and finally eliminates candidate trajectories that violate safe passage requirements by combining constraint execution strategies with safety redundancy requirements.

[0025] The aforementioned personalized trajectory selection and semantic output module, PTS-SL, selects the trajectory that best matches the current driving style and scenario requirements from the candidate trajectories that have undergone constraint screening. It then outputs the final predicted trajectory according to four configurable output modes. The output trajectory undergoes post-processing and smoothing optimization before finally being output to the downstream control module through a preset output interface and communication protocol. For detailed explanations, please refer to the following section.

[0026] In addition, the above modules are as follows: The virtual anchor point generation and adjustment module includes an environmental encoder, a driving style modulator, an aiming point decoder, a noise predictor, and a trajectory distribution penalizer. Among them, The aforementioned environment encoder is used to encode and compress the input environment perception features, converting the multi-source heterogeneous raw perception data into hidden features of a unified dimension, which are then input into subsequent network processing.

[0027] The aforementioned driving style modulator is used to perform dimensionality adaptation and feature modulation on the input driving style embedding, generating a style condition vector that can be injected into the network.

[0028] The aforementioned aiming decoder uses a Transformer decoding structure to progressively generate a temporally arranged candidate estimation sequence.

[0029] The noise predictor described above is used to predict the noise residual at each step during the diffusion sampling process, ensuring the stability of the generated sequence.

[0030] The aforementioned trajectory distribution penalizer is used to constrain and penalize the distribution of the generated candidate estimated sequences, preventing the generated trajectories from exhibiting distribution deviations and ensuring that the generated results conform to the trajectory distribution of the target driving style.

[0031] Those skilled in the art will understand that the above module division is only an exemplary illustration, and in actual implementation, modules can be merged or further split according to requirements.

[0032] Figure 2 The diagram shown is a flowchart of the predicted trajectory generation method provided in an embodiment of this application.

[0033] like Figure 2 As shown, the method for generating the predicted trajectory may include, but is not limited to, the following steps 110 to 130: Step 110: Based on the vehicle's environmental perception information, predict the noise level of the environmental perception information.

[0034] The aforementioned environmental perception information is used to represent at least one of the vehicle's external environmental perception information and real-time vehicle driving status.

[0035] The aforementioned environmental perception information may include, but is not limited to, at least one of obstacle location and road structure. Obstacle information includes both static and dynamic obstacles, and may carry information such as obstacle type, location, speed, and direction of movement. Road structure is used to characterize the topological structure and geometric position of the current vehicle's lane.

[0036] The aforementioned real-time vehicle driving status may include, but is not limited to, at least one of the vehicle's current position, driving speed, and steering angle.

[0037] In some implementations, environmental perception information can be directly output by the driver assistance perception module without the need to introduce additional navigation guide line information as a strong constraint input, thereby reducing the dependence of the trajectory generation process on guide lines.

[0038] Step 120: Based on environmental perception information, remove noise and generate multiple sets of anchor points; multiple sets of anchor points are used to represent multiple predicted trajectories in multiple directions.

[0039] Among them, the aiming point can be a sparse sequence of future path intention points distributed in the lateral space, output by the virtual aiming point generation network. Each point represents the lateral position that the vehicle may occupy at a future time (such as 1s, 2s...5s), which is determined by environmental perception and driving style and does not depend on the guide line.

[0040] Based on the aforementioned environmental perception information, a multi-dimensional decision-making basis is provided for anchor point generation, ensuring that the anchor point generation process does not rely on preset guide lines but is driven solely by the current actual environmental characteristics. Furthermore, by removing noise interference from the environmental perception information, effective environmental features can be retained to stimulate multi-directional trajectory generation intentions, avoiding the problem of trajectory output convergence when guide lines fail, and fundamentally improving the overall robustness of the trajectory prediction system.

[0041] The predicted trajectory is used to represent the guide lines for assisted driving. These guide lines may be, for example, at least one of lane lines and navigation paths.

[0042] Step 130: Determine the final predicted trajectory from multiple predicted trajectories.

[0043] The aforementioned final predicted trajectory represents the final predicted feasible trajectory. Each trajectory is accompanied by a semantic label and probability, making the decision-making process transparent and verifiable, and meeting the mandatory interpretability requirements of ISO (International Organization for Standardization) 26262 functional safety standard, so that it can be used by subsequent L4 level systems.

[0044] In this embodiment, by driving anchor point generation based on environmental perception information, the dependence on guide lines is avoided. Even in scenarios where guide lines fail, candidate predicted trajectories in multiple directions can be output without the problem of trajectory output convergence. This effectively improves the overall robustness of trajectory prediction in assisted driving and ensures the stability and reliability of assisted driving functions.

[0045] Physically constrained trajectory sampling methods in related technologies incorporate vehicle dynamics models and collision detection during the planning phase to perform hard filtering on the sampled trajectories. However, the multimodal nature of these methods relies on random sampling and lacks semantic intent guidance, leading to output trajectories that violate vehicle dynamics constraints, such as excessive lateral acceleration and abrupt curvature changes, posing safety risks.

[0046] To address the technical problems in the aforementioned related technologies, such as multimodal performance relying on random sampling, lack of semantic intent guidance, and output trajectories often violating vehicle dynamics constraints, this application provides a trajectory generation method. This method obtains candidate trajectories through hard constraints of a dynamic sector-shaped feasible region. The dynamic sector-shaped feasible region is determined based on the real-time vehicle driving state and the vehicle's own dynamic parameters from the environmental perception information. This makes the dynamic sector-shaped feasible region a physically feasible constraint, fundamentally preventing the generation of infeasible trajectories that violate dynamics constraints, ensuring that all candidate trajectories are physically feasible, and achieving a safety-first underlying guarantee.

[0047] Figure 3 The diagram shown is a flowchart illustrating the predicted trajectory generation method provided in this application, which includes a dynamic sector feasible region.

[0048] like Figure 3 As shown, step 130 in this trajectory prediction generation method may further include, but is not limited to, the following steps 210 to 230: Step 210: Based on the real-time vehicle driving status in the environmental perception information, determine the dynamic sector feasible region used to constrain candidate points.

[0049] The aiming points mentioned above in this specification are not directly used for vehicle control, but rather serve as the semantic basis for trajectory generation. After the subsequent dynamic sector feasible region is defined, spline interpolation is used to generate the final predicted trajectory.

[0050] In addition, the dynamic sector feasible region DFFR in this embodiment serves as the hard physical boundary of all candidate trajectories, and the dynamic sector feasible region can be adaptively adjusted with vehicle speed.

[0051] The aforementioned dynamic sector-shaped feasible area is adjusted in real time according to the vehicle's current driving status: For example, when the vehicle is traveling at high speed in a straight line, the fan-shaped search range is narrowed to avoid generating lane-changing and turning trajectories with large curvature.

[0052] Example 2: When the vehicle is in a low-speed, turn-preparing condition, the fan-shaped search range is widened to cover possible lateral candidate areas, ensuring that feasible trajectories are not filtered out, thus balancing safety and trajectory coverage.

[0053] The dynamic sector feasible region of the Dynamic Sector Spatiotemporal Hard Constraint Module (DFFR) is defined as follows: The origin is the current position of the center of the rear axle of the vehicle; radius ,in For look-ahead time, such as 3 seconds, At the current vehicle speed, ensure the trajectory does not exceed a reasonable prediction range; angle in The reference angle is, for example, ±30°. The vehicle speed sensitivity coefficient is, for example, 0.01.

[0054] Thus, the higher the vehicle speed, the narrower the fan-shaped area, which aligns with the common sense of human driving that decision-making should be more focused at high speeds. All trajectories generated by VWG (Virtual Waypoint Generator) fall within the Dynamic Fan-Shaped Feasible Area (DFFR); otherwise, if they exceed the DFFR, the trajectory is either discarded or penalized with a penalty, such as +1000.

[0055] Step 220: Based on the constraint of the dynamic sector feasible region as the trajectory generation, the trajectory that does not exceed the dynamic sector feasible region among the multiple predicted trajectories is taken as the candidate trajectory containing the candidate points.

[0056] Here, candidate points represent the possible locations where the vehicle may arrive at multiple predicted times in the future. Connecting all candidate points sequentially yields the complete predicted trajectory. By retaining only the trajectory segments corresponding to candidate points within the dynamic sector range, trajectories exceeding the dynamically feasible range can be directly excluded in advance. Step 230: Based on all the obtained candidate trajectories, and in conjunction with preset trajectory scoring rules, the optimal trajectory is selected as the final output predicted trajectory.

[0057] Step 230: Determine the final predicted trajectory from the candidate trajectories.

[0058] In this embodiment, constraints are pre-defined by a dynamic sector-shaped feasible region to address the technical problem of physically infeasible trajectories. Thus, in the initial stage of trajectory generation, physically infeasible candidate trajectories can be excluded, eliminating the need for dynamic verification after generating all trajectories. This also reduces the consumption of ineffective computational resources and fundamentally avoids the output of dangerous trajectories that violate constraints, thereby improving the safety and operational efficiency of the assisted driving path planning process.

[0059] Furthermore, by embedding physical priors during the generation phase, dangerous trajectories such as lane changes, reverse driving, and high-speed sharp turns are fundamentally avoided, achieving a safe default configuration trajectory basis, which is superior to later filtering or soft constraint methods.

[0060] Combination Figure 2 and Figure 3 As shown, step 110 above can be implemented using at least one of the following optional embodiments: predicting the noise of the environmental perception information based on the vehicle's environmental perception information: In the first optional embodiment of step 110 above, step 310 involves determining environmental characteristics and predicting the noise of the environmental perception information based on the external environmental perception information and the real-time vehicle driving status in the environmental perception information.

[0061] Step 320: Based on the environmental features, the noise is gradually removed by the aiming point decoder to obtain multiple sets of initial aiming point sequences.

[0062] Step 330: Perform output post-processing and smoothing on the multiple sets of initial aiming point sequences to obtain multiple sets of aiming points that meet the requirements of continuous motion and have clear intention semantics.

[0063] Furthermore, as an optional embodiment of this application, the method further includes: the aiming decoder includes a trained noise predictor, which predicts the noise of the environmental perception information based on the vehicle's environmental perception information; the trained noise predictor is obtained through the following steps 410 to 450: Step 410: Obtain the true ideal anchor point sequence in the current vehicle scene and the environmental perception features in the current vehicle scene; the environmental perception features include environmental features and style features.

[0064] Among them, the ideal anchor point sequence is used to represent the standard trajectory without noise interference. It can completely and accurately reflect the vehicle's true driving intention in the current scenario and is the benchmark ground truth data used for annotation during model training.

[0065] For example, the environmental perception feature E is obtained as input, and the final output is N sets of aiming points (N=5), each set being... , =5 (one point every 1 second for the next 5 seconds), for a total of 5×5×2=50 coordinate points.

[0066] in, For the first The framework of the future trajectory corresponding to the driving intention of the group. Used to indicate based on the current environment E The inferred first This provides a set of possible, discretized future driving paths, offering downstream trajectory optimization and feasibility verification modules a set of path point constraints with clear semantics, such as aggressive left lane changing or conservative following, to facilitate trajectory generation and selection. Here, n represents the category index of driving intent. This is used to distinguish between multiple different driving strategies generated simultaneously by the virtual aiming point generation and modulation module. For example, n =1 may indicate that the current lane following operation is enabled. This may indicate a left lane change. This may indicate that a right lane change is being executed. This is a preset hyperparameter representing the number of diverse intent choices the system provides to downstream planning modules at a time. (The above...) This reflects the expected longitudinal distance the vehicle will travel at the end of the T-th second in the future. (The above...) It reflects the lateral offset that the vehicle is expected to achieve at the end of the Tth second in the future. This indicates a rightward shift.

[0067] In summary, the output of the virtual aiming point generation and modulation module is: Each coordinate value. N The sequence of groups together constitutes a predictive trajectory covering a variety of possible future driving intentions, providing a rich and structured selection space for downstream decision-making and planning modules.

[0068] Step 420: Add real noise to the ideal anchor point sequence at each time step to determine the noisy anchor point sequence at each time step.

[0069] The aforementioned "real noise" refers to random noise with a preset distribution added to the ideal anchor point sequence.

[0070] Step 430: Input the noisy aiming point sequence, the environmental features, and the style features into the noise predictor to be trained, so as to output predicted noise.

[0071] Step 440: Update the model parameters of the noise predictor to be trained using backpropagation, and repeat the training iteration until the error between the predicted noise and the real noise decreases to the end of the iteration, thus obtaining the trained noise predictor.

[0072] Step 440 above may include, but is not limited to, obtaining a loss value based on the error between the real noise and the predicted noise; updating the parameters of the noise predictor using the loss value through backpropagation, so that the error between the predicted noise and the real noise decreases to the end of the iteration condition, thus obtaining a trained noise predictor.

[0073] The iteration termination condition may include, but is not limited to, the number of training iterations reaching a preset maximum number of iterations, or the loss value decreasing to below a preset loss threshold. The iteration stops when either of these conditions is met.

[0074] In this embodiment, after training, the trained noise predictor can learn the correspondence between environmental features, style features, and anchor point sequence noise distribution. It can accurately output noise prediction results that conform to scene constraints and driving intentions for different driving scenarios, facilitating the subsequent generation of reasonable anchor point sequences. Furthermore, during the inference phase, the trained noise predictor can gradually obtain aiming point sequences that conform to the current environment and driving intentions from random noise through multi-step denoising, thus completing the aiming point generation process under multimodal conditions. In this way, driving style is embedded into the intention generation layer, i.e., virtual aiming points, realizing that style determines intention, and intention determines trajectory, meeting the personalized needs of assisted driving.

[0075] In some application examples, the network of the aiming point generation and modulation module VWG-SM includes a conditional Transformer and a network based on conditional layer normalization.

[0076] The encoder is used to extract spatial semantics from environmental features E through a 16-layer Transformer encoder using BEV (Bird, bird's-eye view) and Transformer, and output environmental semantic features F_env.

[0077] The driving style modulator is used to map the style feature (Driver Style Embedding, DSE) vector to a set of modulation parameters g=[g_scale, g_shift] through a multilayer perceptron (MLP).

[0078] A network based on conditional layer normalization is used for style modulation, which is performed in the layer normalization (LayerNorm) operation of each layer of the Transformer. This embodiment implements this through a network based on conditional layer normalization, which performs a stylized affine transformation on the feature distribution. The process is as follows: First, the layer input feature x is standardized to eliminate its inherent biases in scale and offset. .

[0079] Subsequently, parameters generated by the driving style modulator were used. and Affine transformation is performed on the standardized features to inject style information into the features. .

[0080] Here, x is the original input activation value of a neuron (or feature vector) in the current Transformer layer, which serves as the output of the previous layer and contains processed environmental information. To obtain features in the current training batch and data space, such as the height / width dimensions of the feature map. The mean value is used to characterize the center position of the feature activation values ​​of this layer and is used for centering. To obtain features within the same statistical range The standard deviation is used to characterize the dispersion of the feature activation values ​​of this layer. It can be scaled to give the data unit variance. These are standardized features with a mean of 0 and a variance of 1, used to stabilize the training process, accelerate convergence, and make the network insensitive to small shifts and scaling of the input. The scaling parameter is learned from the driving style and is used to control the influence of style on the activation intensity of feature neurons. This can amplify the response of the features. This can suppress the response, thereby selectively enhancing or diminishing the importance of certain semantic features based on different driving styles. These different driving styles could be, for example, aggressive or conservative. Semantic features could include attention to distant vehicles. Shift parameter (Bias) is learned from driving style and is used to directly change the activation threshold of neurons, guiding the network to activate different subsets of neurons, i.e., different paths, under different driving styles. y is the final output feature after style modulation, used to preserve the original environmental information. Its distribution is adjusted to a specific pattern defined by the current driving style, thereby achieving style-conditional feature representation.

[0081] In summary, and Together, they perform an affine transformation on the standardized feature distribution, essentially providing the model with a style adjustment method. Through this affine transformation, the same set of network weights can dynamically change the representation of its internal features, thereby activating different neuronal paths under different driving styles and ultimately generating aiming points that conform to a specific style.

[0082] In the embodiments of this application, the style feature DSE vector can be continuously evolved through online learning, such as sliding window updates, thereby sensing changes in driver habits, such as from conservative to aggressive, and automatically adjusting the trajectory generation strategy, thus possessing long-term adaptive capability to achieve personalized assisted driving for each individual.

[0083] A conditional diffusion model in the aiming decoder is used to generate aiming sequence, avoiding the mode collapse problem of GNN / Transformer. The conditional diffusion model may include, but is not limited to, a noise predictor and a trajectory distribution penalizer. The trajectory distribution penalizer (TDP) is not used in subsequent inference.

[0084] Throughout the training process, the noise predictor and trajectory distribution penalty are trained and updated synchronously. During training, for each input condition, a forward process is executed (from adding noise to U-Net predicting noise), and a backward process is executed (sampling multiple paths to reconstruct the complete aiming point trajectory). Simultaneously, the diffusion loss and diversity penalty loss are backpropagated to update the parameters of the noise predictor and trajectory distribution penalty, allowing them to learn accurate denoising while simultaneously learning the characteristics of outputting diverse trajectories. Detailed explanation follows: The aforementioned forward process is used to determine the ideal aiming point sequence. Gaussian noise is gradually added, and after T steps it becomes pure noise.

[0085] The above reverse process is used to gradually denoise a U-Net network based on conditions (environmental feature E and style feature DSE) and reconstruct the aiming point sequence.

[0086] The above diffusion loss function minimizes the mean square error between the predicted noise and the actual noise. .

[0087] in, These are ground truth future waypoint sequences, used as the target for the model to learn during training, and also the starting point for the diffusion process. To sample real noise from a standard Gaussian distribution during the forward process. The random perturbation added to the data at each diffusion step is the target value that the model needs to predict. The cumulative product of the variance schedule for the forward process is a predefined coefficient that monotonically decreases with increasing time step t, used to represent the original signal. In the The remaining weights or signal-to-noise ratio at each step. The closer it is to 1, the more original signal is contained in the noise data at step t. This is a square root operation used to extract the square root of the noise dispatch coefficient to ensure that the way noise is added is consistent with the probability derivation of the diffusion process. This term together constitutes the result of the forward process at step t. Used for scheduling based on noise, to move raw data and noise By mixing with specific weights, the result is obtained at time step Noise data at that time. A denoising U-Net network for parameter θ, also known as a noise predictor, is used to denoise the first... Step noise data Current time step Taking conditional information (F_env, style feature DSE) as input, the noise predictor outputs a predicted value for the noise ε added to the data. In this way, it guesses the true noise as accurately as possible, thus enabling efficient noise removal during the reverse process.

[0088] As an optional embodiment of this application, the trajectory generation model includes a trained trajectory distribution penalizer, and the method further includes: using the trained trajectory distribution penalizer to remove the noise based on the environmental perception information, and generating multiple sets of anchor points; the trained trajectory distribution penalizer is obtained by the following steps 510 to 530: Step 510: Based on the environmental characteristics and driving style characteristics of the vehicle under the current environment, reverse noise reduction is performed on the aiming point sequence with added noise to obtain multiple sets of initial aiming point sequences.

[0089] Step 520: Input the multiple sets of initial anchor point sequences into the trajectory distribution penalizer to be trained, so as to output the distribution likelihood corresponding to each set of initial aiming point sequences.

[0090] Step 530: Determine a diversity loss term based on the distribution likelihood; the diversity loss term is used to increase the diversity of the multiple predicted trajectories under the same input conditions; the diversity loss term is negatively correlated with multiple initial aiming point sequences whose differences are less than a preset difference.

[0091] The preset difference is used to represent the difference in distribution distance between different initial anchor point sequences in the trajectory feature space. When the distribution difference of multiple initial anchor point sequences is less than the preset difference, it indicates that the different initial anchor point sequences are relatively similar, also known as excessive trajectory homogenization. At this time, the value of the diversity loss term increases, which forms a reverse penalty on the model parameters, pushing the model to adjust the generation direction and output multiple sets of anchor point sequences with greater differences.

[0092] Step 540: Using backpropagation, update the model parameters of the trajectory distribution penalty device to be trained, and repeat the training iteration until the loss of the trajectory distribution penalty device to be trained meets the iteration termination condition, thereby obtaining the trained trajectory distribution penalty device.

[0093] In the embodiments of this application, by introducing a diversity loss term to penalize the trajectory homogenization problem, it is possible to ensure that the anchor point generation covers different potential driving intentions while avoiding multiple sets of predicted trajectories being too concentrated, thereby improving the diversity and coverage of predicted trajectories and better adapting to the uncertainty of the driver's potential driving intentions in assisted driving scenarios.

[0094] As another optional embodiment of this application, the trajectory distribution penalty device also adjusts the calculation weight of the current distribution likelihood by combining the distribution information of historically generated trajectories. For distribution intervals that recur in historical generation, the likelihood score of the anchor point sequence within the corresponding interval is reduced, further ensuring the overall diversity of predicted trajectories generated in different rounds. This is done when calculating the total loss. At that time, the diversity loss term With denoising and reconstruction loss term The weighted fusion total loss can be expressed as: ;in, , These are configurable hyperparameters, and their weight ratios can be adjusted according to the requirements for trajectory accuracy and diversity in actual scenarios.

[0095] In other training embodiments of the trajectory distribution penalizer, with fixed input environmental features and driving style conditions, the trajectory distribution penalizer to be trained generates multiple predicted trajectories through random sampling in a diffusion process. The variance of the aiming point coordinates of the predicted trajector at each time step is determined to measure the dispersion between trajectories. Based on the dispersion, a diversity loss term is determined; this diversity loss term is used to increase the diversity of the multiple predicted trajectories under the same input conditions; the diversity loss term is positively correlated with the dispersion. Backpropagation is used to update the model parameters of the trajectory distribution penalizer to be trained, and training iterations are repeated until the loss of the trajectory distribution penalizer to be trained meets the iteration termination condition, resulting in a trained trajectory distribution penalizer. Thus, in the training of the trajectory distribution penalizer, a penalty is imposed to encourage diversity.

[0096] To avoid generating overly similar trajectories, a diversity penalty term (also known as a diversity loss term) is introduced during the trajectory distribution penalty selection training process: ,in It is variance. =0.001, to prevent numerical overflow.

[0097] Total loss: ,in =0.1.

[0098] in, In a single training batch, for the same input conditions—namely, the same environmental feature F_env and the same driving style embedding style feature DSE—the model generates N distinct future trajectories at time step t through random sampling, resulting in the waypoint state. This state contains at least... Coordinates. L_div is the diversity loss term used to encourage the generation of results.

[0099] First, for each time step t, calculate the variance (Var) of the state values ​​of the N trajectories at that time step, which measures the dispersion of the trajectories at that moment. Then, sum the variances of all time steps to obtain an overall diversity measure. Finally, take the negative logarithm of the summation. Thus, when the sum of variances is small, it indicates that the trajectories are similar, and the penalty (smallest value of -log) will be large; when the sum of variances is large, it indicates that the trajectories are diverse, and the penalty (largest value of -log) will be small. Therefore, minimizing L_div is maximizing the diversity of generated trajectories.

[0100] L_diffusion is the main loss function of the diffusion model, which requires the noise prediction network to predict the noise added to the real data during the forward pass as accurately as possible. This is the weighting coefficient for the diversity loss term, used to control the relative importance of the diversity penalty term in the total loss. This application's embodiments aim to generate accurate and conditional trajectories primarily through the model, which is derived from... Driven by, and supplemented by, appropriate diversity incentives, diversity incentives are... and The driver. The value is a hyperparameter that needs to be adjusted based on the experimental results.

[0101] Subsequently, cubic spline interpolation is performed on the generated N trajectories to ensure... Continuity (position, velocity, and acceleration are continuous). Additionally, the lateral acceleration variance for each trajectory is calculated as a smoothness indicator. This is for output post-processing and smoothing.

[0102] In this embodiment, the problem of pattern collapse is solved by using a lateral distribution penalty and diversity optimization mechanism.

[0103] After filtering by the Dynamic Sector Spatiotemporal Hard Constraint (DFFR) module, the lateral position distribution of the remaining M candidate trajectories at key time points is calculated. The lateral distribution variance is defined. =Var{y_t}. Key time points include, for example, 1 second and 2 seconds.

[0104] During the training of the trajectory distribution penalty device, a distribution penalty term is used to transform the lateral spatial distribution breadth of the trajectory into an optimizable signal that encourages diversity, based on the concept of entropy increase in information theory. The distribution penalty term is as follows: ε is a small constant to prevent numerical instability. This penalty term is added to the total loss function and jointly optimized with the generation loss of VWG, incentivizing the model to generate trajectories with a more dispersed lateral distribution, thus avoiding all trajectories clustering in the same lane.

[0105] In this embodiment, diverse intents are generated through VWG, and TDP prevents them from crowding together due to optimization convergence, ensuring that diversity is real and meaningful, effectively combating pattern collapse.

[0106] As an optional embodiment of this application, determining the user's driving style characteristics based on the external environment perception information and real-time vehicle driving status in the environmental perception information includes inputting the external environment perception information and real-time vehicle driving status into the driving style embedding extraction module to output the user's driving style characteristics; the driving style embedding extraction module is trained using a style label dataset.

[0107] In this embodiment, the driving style features extracted by the driving style embedding extraction module can be embedded into the trajectory generation network, so that the generated trajectory can match the current user's driving habits and meet the personalized needs of different users for trajectory comfort and aggressiveness.

[0108] Step 110 above may further include, but is not limited to, extracting environmental features from the external environment perception information and the real-time vehicle driving status in the vehicle's environmental perception information.

[0109] Accordingly, the method further includes: inputting the environmental features and the driving style features into the aiming point decoder through conditional layer normalization.

[0110] Furthermore, the driving style embedding extraction module may include, but is not limited to, a preprocessing module, an intermediate layer, and a time attention mechanism.

[0111] Step 610: Based on the external environment perception information and real-time vehicle driving status in the environmental perception information, the preprocessing module performs dimensionality upscaling mapping to obtain the features of each time step.

[0112] Specifically, based on the external environment perception information and real-time vehicle driving status in the environmental perception information, an original time-series driving signal is generated. This original time-series driving signal is then subjected to dimensionality up-mapping to obtain the features of each time step.

[0113] Step 620: Based on the features of each time step output by the preprocessing module, the intermediate layer fuses the context information of adjacent time steps and extracts the hidden state of each time step of the fused context information.

[0114] Step 630: The weights of each time step are learned through the time attention mechanism, and the hidden states of each time step are weighted and summed with the weights of the corresponding time steps to generate a third-style embedding vector.

[0115] The third style embedding vector obtained above can represent the current user's driving style, avoiding noise interference and feature redundancy caused by directly using the original signal, and improving the style matching degree of subsequent trajectory generation.

[0116] In this embodiment, the preprocessing module is used to obtain the features of each time step, which can retain the change information of the user's driving operation under different time dimensions. By using the time attention mechanism to highlight the influence of key time steps on the overall driving style through weight allocation, style features that can reflect the user's long-term driving habits can be extracted, avoiding feature extraction errors caused by fluctuations in single time step information and improving the reliability of driving style features.

[0117] Combination Figure 1 As shown, the driving style embedding extraction module DSEE can be used as a neural encoder representing driving style.

[0118] 1. Input data stream and sampling mechanism Input source: High-frequency (≥20Hz) driving status signals output in real time by the vehicle domain controller (ADC), including: longitudinal acceleration. lateral acceleration Steering wheel angle Steering wheel angle change rate Longitudinal distance from the vehicle in front, driving style embedded vector Relative speed to the vehicle in front Did a lane-changing incident occur? (Based on lane line detection + lateral displacement threshold ≥ 0.8m judgment), vehicle speed . The sampling time interval is the time difference between two consecutive data acquisitions by the sensor. The current steering wheel angle. This represents the steering wheel angle at the previous moment.

[0119] The aforementioned input signal comes from the vehicle domain controller (ADC), in order to The following real-time driving state variables are collected at a frequency: longitudinal acceleration lateral acceleration Steering wheel angle Steering wheel angle change rate (Calculated based on the difference between adjacent sampling points), longitudinal distance to the vehicle in front d、 Relative speed to the vehicle in front 、 Lane change event sign (Determined by lane line detection and lateral displacement threshold ≥ 0.8m), vehicle speed All variables are time series data, with sampling intervals of [missing information]. This is a fixed value. The aforementioned time window refers to a sliding window. seconds, sampling frequency Therefore, each sample is 3D time-series vector (600 time steps × 8 features).

[0120] 2. The network structure and feature encoding are as follows: The encoder adopts a bidirectional GRU (Bi-GRU) and temporal attention pooling structure to capture long-term dependencies in driving behavior and focus on key operational segments. The input layer receives sequence features generated by the preprocessing module, with dimensions of 600 (time step) × 256 (feature dimension).

[0121] The Bi-GRU layer processes the input sequence, and at each time step outputs a hidden state that integrates contextual information. A temporal attention mechanism is used to determine the hidden state at each time step using the following formula. attention weights The formula is: .in, .in, It is a learnable parameter matrix with dimensions 256×256. and They represent the first The score of the hidden state at the first time step and the score of the second time step. The score of the hidden state at each time step reflects the importance of the corresponding time step to the overall driving style.

[0122] Driving style embedding generation, used for the final driving style embedding vector z, is obtained by weighted summation of the hidden states over all time steps: The final output is a low-dimensional, highly abstract 32-dimensional driving style embedding vector z, which is used to represent the driver's unique style.

[0123] in, The attention weight matrix is ​​a learnable matrix used to represent a score for evaluating the importance of the hidden state at each time step. The model learns the matrix through training to automatically identify and amplify key driving segments, such as rapid acceleration and emergency lane changes, that contribute most to differentiating driving styles. Let be the hidden state at time step i, used by the Bi-GRU network to generalize the input sequence at time step i, encoding the driving behavior context information up to the current moment. 'i' is the index of the current time step, used to traverse each point in the sequence, such as second 1, second 2, etc. 'j' is the summation loop index, used in the normalized denominator of the attention weights (...). This is used to represent all possible time steps in a traversal sequence to ensure that the sum of all weights is 1.

[0124] In this embodiment of the application, in order to meet the style representation needs of different drivers, the 32-dimensional driving style embedding vector can present the clustering effect of different driving behaviors in a low-dimensional space. The feature vectors of aggressive driving style and conservative driving style will form a distinguishable distribution in the space, and the autonomous learning and extraction of features can be completed without manual labeling of driving style.

[0125] As an optional embodiment of this application, the method may also include, but is not limited to, the following steps 610 to 650 to obtain the style tag dataset: Step 610: Obtain the style embedding vectors corresponding to all driving vehicle segments for each driver, perform three-dimensional scoring and annotation, and obtain the three-dimensional scoring results and basic style annotation information.

[0126] The three-dimensional scoring results mentioned above can refer to the following three dimensions: Aggressiveness, Smoothness, and Proactivity. Aggressiveness is measured by lane-changing frequency, following distance, and peak acceleration. Smoothness is measured by lateral acceleration variance and steering wheel vibration frequency. Proactivity is measured by early deceleration, early lane change, and TTC response delay.

[0127] Based on this, the three-dimensional scoring and annotation of the style embedding vectors corresponding to all driving vehicle segments for each driver in step 610 above may include, but are not limited to: a. Selecting 1000 drivers from real road survey data, each providing 5000 30-second segments. b. Having human annotators give each driver a three-dimensional style score, for example, from 1 to 5 points.

[0128] Step 620: Based on the three-dimensional scoring results and the basic style annotation information, generate a first style embedding vector.

[0129] Step 630: Cluster the first style embedding vector to obtain style category pseudo-labels and second style embedding vectors.

[0130] The clustering methods include any combination of K-Means clustering, hierarchical clustering, spectral clustering, etc. Once clustering is complete, the cluster center of each category is retained as the second style embedding vector for the corresponding style category. Simultaneously, a pseudo-label for the corresponding style category is assigned to each first style embedding vector, thus completing the grouping and classification of style features and providing a structured style feature foundation for subsequent style modulation.

[0131] The clustering of the first style embedding vector in step 630 above may further include: K-Means clustering, which clusters the first style embedding vector.

[0132] For example, when K=5, K-Means clustering is used to cluster the 32-dimensional style feature DSE vectors, resulting in 5 style categories: S1 is Aggressive, S2 is Cautious, S3 is Calm, S4 is Lagging, and S5 is Emotional. Thus, the first style embedding vector is assigned to each of these 5 style categories, with each style category corresponding to a 5-dimensional one-hot encoded style category pseudo-label. Simultaneously, the 5 cluster centers are extracted as the second style embedding vectors for their respective style categories.

[0133] Step 640: Input the style category pseudo-label and the second style embedding vector into the driving style embedding extraction module, and output the optimized third style embedding vector.

[0134] Among them, contrastive loss is used to constrain the optimization parameters of the driving style embedding extraction module, so that the second style embedding vector is clustered in the same category and separated in the opposite category.

[0135] For example, a contrastive loss model is used to optimize the model, bringing similar (same driving style) samples closer together and distancing dissimilar (different driving style) samples further apart in the embedding space. The modified loss function is defined as follows:

[0136] Among them, the above For the contrastive loss function, a smaller value indicates that the embedding space learned by the model better matches the expectation of similar classes and dissimilar classes. 'a' is the index of the first sample in the sample pair. It is used to iterate through each sample in the batch (a∈[1, N]). 'b' is the index of the second sample in the sample pair, used to iterate through each sample in the batch and form a pair with sample 'a'. (∈[1, N]). N is the total number of samples in a training batch. The driving style category label for the a-th sample is used to indicate the specific driving style of the sample data, such as aggressive or mild driving, for a particular driving trip. For the first Driving style category labels for each sample. Let be the driving style embedding vector (style feature DSE Vector) of the a-th sample. It is a 32-dimensional feature vector generated by the encoder network for the a-th sample and is a numerical representation of the driving style of that sample. For the first The driving style embedding vector of each sample. I(·) is the indicator function, which outputs 1 when the condition in parentheses is true, and 0 otherwise. =1 means that sample a and sample a = 1 if and only if sample a and sample a = 1. They belong to the same driving style. I( ) =1 means if and only if sample a and sample These belong to different driving styles. ||·|| is the L2 norm (Euclidean distance) of the vector. Here || The || parameter is used to determine the actual distance between two embedding vectors in space. m is the margin threshold, a hyperparameter greater than 0; here, m = 1.5, which specifies the minimum distance that out-of-class samples must maintain. A loss value is generated only when the distance between out-of-class samples is less than this threshold.

[0137] Step 650: Based on the third style embedding vector, re-cluster and update the pseudo-labels to obtain the style label dataset.

[0138] The third style embedding vector has a higher dimension than both the second and first style embedding vectors, and the second style embedding vector has a higher dimension than the first style embedding vector. Specifically, the third style embedding vector has a dimension of 20 or higher, and the second style embedding vector has a dimension of 8 or lower. For example, the high-dimensional vector is a vector with 32 dimensions, and the low-dimensional vector is a vector with 5 dimensions.

[0139] In this embodiment, through multiple iterations of annotation, clustering, contrastive learning optimization, and re-clustering, a style label dataset with higher discriminative power and better conformity to the distribution of real driving styles can be obtained. This provides reliable training and supervision signals for the subsequent aiming point generation network of style modulation, ensuring that the network can learn effective features of different driving styles and solving the problem of mismatch between intent generation and driver style in trajectory prediction.

[0140] As an optional embodiment of this application, the above method further includes: performing time-series smoothing optimization on the current driver's driving style results to obtain new driving style results: according to the new driving style results within the current time window, by using an exponentially weighted moving average and fusing historically accumulated style information, the current driver's driving style is obtained to characterize the driver's long-term and stable driving habits.

[0141] The engineering implementation and online updates in this application embodiment are as follows: The deployment model is that the driving style embedded extraction module DSEE is deployed on edge computing units such as NVIDIA Orin, with an inference latency of ≤15ms.

[0142] Online Adaptation: The style feature DSE vector is updated every 5 minutes, using an exponentially weighted moving average (EWMA): Style feature DSE_t = λ Style feature DSE_{t-1}+(1-λ) Style feature DSE_new, where λ=0.95, t The time step index indicates the number of times the system is running. t The update time (in 5-minute increments) is used to identify the temporal position of the current state, reflecting the system's continuous memory and evolution tracking of historical states; λ =0.95 is the smoothing factor (or attenuation coefficient), representing the system's memory weights for the historical style feature DSE vector; its physical meaning is that in the current estimate, 95% of the weight comes from the historical cumulative state, and 5% of the weight is assigned to the latest observation. This value ensures that the system has sufficient responsiveness to environmental changes while effectively suppressing the influence of sensor noise and transient disturbances, achieving an adaptive balance of "stability with change".

[0143] If a sensor malfunctions, such as loss of steering wheel signal, it automatically switches to the default stable mode (S3) as a degradation strategy to provide a fault tolerance mechanism.

[0144] Combination Figure 2 As shown, as the first optional embodiment of step 210 in this application, step 210 may include, but is not limited to, step 710. Correspondingly, step 220 may include, but is not limited to, step 720: Step 710: Delineate the dynamic sector feasible area based on the vehicle's current driving state and its own dynamic parameters.

[0145] Step 720: Based on the constraints of the dynamic sector feasible region as trajectory generation, delete all trajectory points of the predicted trajectory that exceed the maximum lateral acceleration and allowable curvature range of the vehicle, and obtain the trajectory of the predicted trajectory that does not exceed the dynamic sector feasible region as the candidate trajectory. The boundary of the dynamic sector feasible region is determined by the maximum lateral acceleration and allowable curvature corresponding to the current vehicle speed.

[0146] In this embodiment of the application, applying dynamic sector hard constraints before generating the predicted trajectory can guarantee 100% physical feasibility of the output trajectory without the need for an additional safety verification module, thereby effectively improving system efficiency and reliability.

[0147] As a second optional embodiment of step 210 of this application, step 810 uses the current coordinates of the rear axle of the vehicle as the origin of the sector.

[0148] Step 820: Based on the instantaneous speed of the vehicle and the configurable look-ahead time, obtain the dynamic radius of the sector and the total angle of the sector.

[0149] Step 830: Adjust the sector origin, sector dynamic radius and sector total angle to obtain a dynamic sector feasible region for constraining candidate points.

[0150] There are several ways to adjust the above step 830 based on the sector origin, the sector dynamic radius, and the sector total angle. In the first adjustment method, the adjustment is made according to the real-time position based on the sector origin, the sector dynamic radius, and the sector total angle. In the second adjustment method, it may include, but is not limited to, making real-time adjustments based on the sector origin, the sector dynamic radius, and the sector total angle at various time points.

[0151] In this embodiment, by setting the current coordinates of the vehicle's rear axle as the origin of the sector, and combining the instantaneous speed and look-ahead time to determine the dynamic parameters of the sector, and then adjusting them in real time according to different scenarios to obtain the dynamic sector feasible area, the constraint range can be made to better fit the current actual driving state of the vehicle. While ensuring the physical feasibility of the trajectory, it avoids the conservative problem of trajectory planning caused by the fixed constraint range, does not waste feasible road space, and improves the flexibility and planning efficiency of trajectory generation.

[0152] As an optional embodiment of this application, the above method may also include, but is not limited to: when a collision risk is detected in the vehicle, shortening the look-ahead time, reducing the dynamic radius of the sector, and narrowing the total angle of the sector; Accordingly, the above-mentioned adjustment based on the origin of the sector, the dynamic radius of the sector, and the total angle of the sector to obtain the dynamic sector feasible region for constraining candidate points may include, but is not limited to: generating the dynamic sector feasible region according to the origin of the sector, the narrowed total angle of the sector, and the reduced dynamic radius of the sector.

[0153] In this embodiment, by dynamically adjusting the sector parameters according to the collision risk, the range of the feasible area can be quickly narrowed, and candidate trajectory points with risks can be removed more quickly. This improves the safety of vehicle driving without sacrificing the flexibility of normal driving planning, and does not require an additional large amount of computation, thus not slowing down the overall trajectory planning response speed.

[0154] As an optional embodiment of this application, the step of using the dynamic sector feasible region as a constraint for trajectory generation, and selecting the trajectories among the multiple predicted trajectories that do not exceed the dynamic sector feasible region as candidate trajectories containing the candidate points, may include, but is not limited to, steps 910 to 930: Step 910: For each candidate point of the multiple candidate trajectories, determine whether the distance from the current position is not greater than the dynamic sector feasible area, and whether the vehicle's self-heading angle deviation is not greater than half of the angle range at the current moment.

[0155] Step 920: If the conditions are not met, discard the unqualified trajectory.

[0156] Step 930: If the conditions are met, retain the qualified trajectory as a candidate trajectory.

[0157] In this embodiment of the application, by performing dual constraint judgments on the range and angle of each candidate point, candidate trajectories that do not meet the safe driving range can be quickly removed. This ensures that the final retained candidate trajectories are all within the safe dynamic sector feasible area. Furthermore, because the logic of point-by-point judgment is simple and direct, it does not add too much computational burden to the trajectory sampling process, thus maintaining the efficient operation of the trajectory generation process and ensuring the response speed of the vehicle trajectory planning.

[0158] As an optional embodiment of this application, the method is applied to a trajectory generation model; the method further includes: during the training of the trajectory generation model, adding a corresponding penalty term to the candidate trajectories that are close to the boundary; the penalty term is used to learn to avoid the candidate trajectories that are close to the boundary during the training of the trajectory generation model, and the penalty term is negatively correlated with the distance between the candidate trajectory and the boundary of the dynamic sector feasible region.

[0159] Among them, the penalty intensity is determined based on the distance between the candidate trajectory and the boundary of the dynamic sector feasible region. The closer the candidate trajectory is to the boundary, the higher the penalty intensity of the penalty.

[0160] In this embodiment, when calculating the training loss, an additional loss value is added to these types of trajectories. This allows the model to learn during backpropagation parameter updates that such boundary-hugging trajectories are not good enough, thereby adjusting the parameters and reducing the probability of generating such trajectories subsequently. Thus, when an obstacle is detected within the dynamic sector feasible region, a safety policy is directly triggered, and the candidate trajectory is no longer output using the dynamic sector feasible region.

[0161] Combination Figure 1 As shown, the following dynamic sector spatiotemporal hard constraint module DFFR is used to ensure the insurmountable boundary of vehicle physical safety.

[0162] 1. The geometric definition of a sector is as follows: The origin refers to the Cartesian coordinate of the center of the rear axle of the vehicle at the current moment. .

[0163] radius ,in =3.0s (configurable); where, Let t be the maximum radius of the safe area that the vehicle can cover at the current time t. Let be the instantaneous velocity of the vehicle at time t. ). It is a dynamic variable that allows the size of the safe zone to adaptively adjust with changes in vehicle speed. Safety Foresight Time is the future time span used to ensure safety. It is defined as the maximum distance that can be reached in the future at the current instantaneous speed, serving as the upper limit of the safety radius.

[0164] Angle range ,in =30° is the basic angle; =0.01s / m is the proportionality coefficient; when At (108km / h), ; Let be the total angle of the sector region at time t, representing the lateral width of the safe zone. Higher speeds restrict vehicle maneuverability and reduce lateral movement capability, thus requiring narrowing of the feasible region to avoid overly aggressive or infeasible lateral planning. t is a time variable representing the distance from the current time. The calculation begins at a future point in time.

[0165] 2. The constraint determination algorithm is as follows: for candidate points on the candidate trajectory Check point by point whether the distance from the candidate point to the origin is less than or equal to the radius at the current time. : Candidate points whose deviation from the vehicle's heading angle is ≤ half of the current angle range: .in, For the vehicle's heading angle , These are the origin coordinates, defined as the Cartesian coordinates of the rear axle center of the vehicle at the current moment. ; For the vehicle at the current moment The yaw angle is used to define the vehicle's longitudinal axis direction relative to the global coordinate system. The included angle of the axis is the direction of the center line of the sector.

[0166] in, For the trajectory point Corresponding future moment The total angle of the sector region; This represents the future timestamp corresponding to trajectory point j. For example, if the trajectory time interval is 0.1s, then the j=5th trajectory point corresponds to time t. j =0.5s. j is the trajectory point index, used to traverse each trajectory point j on the trajectory from 1 to T.

[0167] 3. If any candidate point violates the above constraints, that candidate point is discarded directly (probability = 0). If it is close to the boundary, for example, the distance is <0.5m, a penalty term is added to the trajectory loss. Where α=1000, this ensures that the trajectory generation model actively avoids boundaries during training. This represents the distance from the candidate trajectory point to the boundary of the dynamic sector feasible region.

[0168] 4. If an emergency braking vehicle is detected ahead, for example... The system automatically shortens the T_foresee to 2.0s, narrows the fan-shaped area, and avoids over-prediction to achieve dynamic threshold adjustment. If the LiDAR detects an obstacle within the fan-shaped area, even if the predicted trajectory is geometrically compliant, a safety priority mode is triggered, forcibly selecting a deceleration trajectory to achieve multi-sensor fusion verification.

[0169] Trajectory prediction methods in related technologies include deterministic trajectory generation based on guideline extrapolation and multimodal sampling methods. Multimodal refers to the model's ability to output multiple candidate trajectories with different directions and intentions, allowing for selection based on different potential driving behaviors.

[0170] In response, deterministic trajectory generation based on guideline extrapolation can only generate a single standardized trajectory and cannot distinguish individual differences; multimodal sampling methods lack reasonable constraints, and the generated trajectories themselves are unreasonable trajectories that do not conform to actual traffic rules, and cannot match the driving styles of different drivers.

[0171] In summary, the trajectory prediction methods in related technologies cannot meet the individual differences among drivers. Therefore, regardless of whether the driver is aggressive, conservative, or moderate, a standardized trajectory is output, leading to a fragmented passenger experience. For example, aggressive drivers may find the system too conservative, while conservative drivers may find it too aggressive. This one-size-fits-all planning strategy significantly reduces user trust and acceptance of driver assistance systems.

[0172] To address the technical problem of significantly reducing user trust and acceptance of driver assistance systems in the aforementioned related technologies, this application provides a trajectory generation method. In this embodiment, the user's driving style characteristics and environmental characteristics are used, with the driving style characteristics reflecting the user's consciousness level. Therefore, multiple predicted trajectories conforming to the driving style characteristics are determined from the consciousness level. These multiple predicted trajectories stimulate reasonable multimodal generation, thereby stimulating the diversity of predicted trajectories, completely eliminating dependence on guide lines and avoiding random perturbations in the trajectory space. Thus, due to the existence of driving style characteristics (DSE), multimodal generation has a personalized semantic basis, not only solving the fundamental problem of having no intention to rely on when guide lines fail, but also ensuring physical feasibility and enabling personalized adaptation to driver style to generate personalized trajectories, thereby fulfilling safety, diversity, and comfort requirements.

[0173] In a second optional embodiment of step 110 above, the user's driving style characteristics, environmental characteristics, and noise of the environmental perception information are determined based on the external environmental perception information and the real-time vehicle driving status in the environmental perception information. Accordingly, the step of removing noise and generating multiple sets of anchor points based on the environmental perception information may include removing noise and generating multiple sets of aiming points based on the user's driving style characteristics and environmental characteristics. Determining the final predicted trajectory from the candidate trajectories may include the following two steps: The first step is to comprehensively score each candidate trajectory based on safety, driving style matching, and comfort to obtain a score for each candidate trajectory.

[0174] The second step is to determine the final personalized trajectory from the candidate trajectories based on their scores.

[0175] In this embodiment, by pre-removing noise, interference from environmental perception errors on driving style judgment and anchor point generation can be avoided, reducing the risk of trajectory deviation caused by abnormal perception data. When multiple sets of anchor points generated after noise removal still show that all candidate trajectory scores are lower than the preset safety threshold, a safety redundancy strategy will be automatically triggered: the system switches to a basic conservative trajectory mode. The trajectory generated in this mode is based on the minimum safety requirements of the current road traffic rules and does not match the user's personalized driving style, prioritizing the basic safety of the driving process. At the same time, the in-vehicle reminder module issues a downgrade prompt to the user's driver assistance system, reminding the user to prepare for takeover and avoiding safety risks caused by the failure of personalized trajectory generation.

[0176] As an optional embodiment of this application, the step of removing noise and generating multiple aiming points based on the user's driving style characteristics and environmental characteristics includes the following steps 1 and 2: Step 1: Based on the environmental features and driving style features, the noise is gradually removed by the aiming point decoder to obtain multiple sets of initial aiming point sequences.

[0177] Among them, environmental features may include, but are not limited to, at least one of the following: vehicle speed, vehicle acceleration, obstacle position and motion information, and lane line information.

[0178] The driving style feature is a DSE feature vector extracted from the user's historical driving data, which can represent the user's personalized preferences for aggressiveness, following distance, and lane change timing during driving.

[0179] The first step mentioned above may further include, but is not limited to: predicting the position coordinates of each time step point by point using the joint encoding results of the aiming point decoder based on features, correcting the position deviation introduced by the initial noise in each prediction step, and finally obtaining multiple sets of initial aiming point sequences containing different driving intentions.

[0180] Step 2 involves post-processing and smoothing the multiple sets of initial aiming point sequences to obtain multiple sets of aiming points that meet the requirements of continuous motion and have clear intentional semantics.

[0181] In this embodiment, after inputting environmental features and driving style features into the aiming point decoder, multiple sets of initial aiming point sequences containing different driving intentions can be obtained through the aiming point decoder. Then, abnormal points that do not conform to vehicle kinematic constraints are filtered out through post-processing, and then multiple sets of valid aiming points are obtained through interpolation and smoothing. Each set of aiming points corresponds to a candidate driving trajectory, providing diverse candidate trajectories for subsequent personalized trajectory selection.

[0182] This application can employ an aiming point generation network to remove noise and generate multiple sets of aiming points based on the user's driving style characteristics and environmental characteristics. In this way, the aiming point generation network can solve the problem of missing intent generation.

[0183] Specifically, one embodiment of this application uses a Virtual Waypoint Generator (VWG) network, which combines environmental context with Driver Style Embedding (DSE) vectors, and outputs N sets of laterally distributed waypoints. The environmental context input to the waypoint generation network is at least one of the following: image, point cloud, high-definition map, and historical trajectory. The number of N sets is, for example, 5 sets.

[0184] The style feature DSE vector is generated by lightweight GRU encoding of the driving behavior sequence (acceleration, steering angle change rate, TTC, lane change frequency, etc.) over the past 30 seconds, representing the driver's latent characteristics such as aggression, smoothness, and lane-changing tendency. Among them, the driving behavior sequence includes at least one of acceleration, steering angle change rate, TTC, and lane-changing frequency.

[0185] The VWG network employs Conditional Layer Normalization (DSE) to inject style features (DSE) into each layer of feature extraction, thereby achieving dynamic modulation of the spatial distribution of style on aiming points.

[0186] For example, aggressive drivers, correspondingly, focus their aim on the left lane, thus inducing a strategy of early lane change and overtaking; Conservative drivers, correspondingly, keep their aiming point stable in front of the current lane, thus employing a long-distance following strategy; For drivers with a stable driving style, the system generates slight leftward or rightward adjustments to the aiming point to avoid drastic maneuvers.

[0187] As an optional embodiment of this application, the step of comprehensively scoring each candidate trajectory according to safety, driving style matching degree and comfort to obtain the score of each candidate trajectory includes weighting each candidate trajectory according to safety score and safety weight, trajectory smoothness and comfort weight, and driving style matching degree and driving style matching weight to obtain the score of each candidate trajectory.

[0188] Accordingly, determining the final personalized trajectory from the candidate trajectories based on their scores includes selecting the candidate trajectory with the highest score from the scores of each candidate trajectory as the personalized trajectory.

[0189] The driving style matching degree is determined by the degree of matching and the weight of driving style matching. The degree of matching is determined as follows: the data features of different driving style trajectories are input into the discriminator to output the degree of matching between any driving style trajectory and the driving style. The degree of matching is used to indicate whether the trajectory conforms to the user's long-term preferences. The comfort level is determined by a combination of trajectory smoothness and comfort weights; the trajectory smoothness is the reciprocal of the variance of the lateral acceleration of the trajectory and is normally correlated with the comfort weights. The safety level is determined by a safety score and a safety weight; the safety score is positively correlated with the distance between the vehicle and the obstacle and negatively correlated with the collision risk.

[0190] In this embodiment, by weighting the candidate trajectories according to their safety score and safety weight, trajectory smoothness and comfort weight, and driving style matching degree and driving style matching weight, a comprehensive score is obtained. The candidate trajectory with the highest score is then selected as the final personalized prediction trajectory. This can ensure the driving safety and ride comfort of the trajectory while meeting the user's personalized driving preferences, and avoid decision-making bias caused by a single evaluation dimension.

[0191] This application's embodiments address the issue of disjointed interpretation styles by employing a personalized trajectory filtering and semantic output mechanism. The aforementioned personalized trajectory is the final output of a set of candidate trajectories with probabilities, each trajectory accompanied by a semantic label, such as 60% for steady following, 30% for preparing to change lanes left, and 10% for emergency deceleration.

[0192] The filtering strategy supports the following four modes: 1. The most likely trajectory is the one with the highest probability. 2. The safest trajectory is the one with the smallest average distance from the disturbance guide line; 3. The most optimistic trajectory is the one that is closest to the original clean leading line; 4. Calculate the style preference score based on the style feature DSE vector, and prioritize the selection of the strategy trajectory that matches the driving style. For example, if the style feature DSE is high aggressive, then the lane-changing trajectory is selected as the personalized trajectory.

[0193] Compared to the standardized trajectories output by related technologies, they cannot adapt to the long-term behavioral preferences of different drivers, such as aggressive lane changing and conservative following, resulting in low comfort and weak user trust.

[0194] Compared to related technologies, the embodiments of this application achieve personalized trajectory output for each individual by constructing a deep coupling mechanism between driving style embedding and trajectory generation. This allows for the selection of personalized trajectories from a variety of options, making the final output not only safe and diverse but also in line with user expectations, thereby improving ride comfort and system trust.

[0195] When candidate trajectories with the same comprehensive score are present, the candidate trajectory with the higher safety score is selected as the final trajectory. This further strengthens the priority of driving safety, completes the safety redundancy backup, avoids unreasonable trajectory selection results due to the scoring mechanism being equal, and ensures that the final output trajectory meets the basic safety requirements in any scenario.

[0196] As an optional embodiment of this application, the method further includes: sending the final predicted trajectory to a vehicle controller, so that the vehicle controller generates control commands for the vehicle based on the final predicted trajectory and the real-time vehicle driving status; the control commands are used to control the vehicle to execute according to the final predicted trajectory. Here, the vehicle controller refers to an electronic control unit for the lateral and longitudinal motion control of the vehicle. After receiving the final predicted trajectory data, the vehicle controller obtains steering, driving, and braking control commands that conform to vehicle dynamics constraints based on the current vehicle driving status, such as wheel speed, turning angle, and acceleration, ensuring that the vehicle can smoothly follow the planned personalized trajectory.

[0197] Furthermore, the aforementioned vehicle controller may include at least one of a body controller, a domain controller, and a model predictive controller (MPC).

[0198] In this embodiment, the final predicted trajectory is a personalized trajectory that meets the requirements of safety, driving style matching, and comfort. This personalized trajectory is then given to the vehicle controller for control. This allows for personalized control based on the differences in driving habits of different drivers, thereby avoiding the problem that the generated trajectory deviates too much from the driver's daily driving habits, which could lead to the driver's distrust of the driver assistance system or even take over the operation. At the same time, it can ensure the safety and smoothness of the output trajectory.

[0199] Combination Figure 1 As shown, the Personalized Trajectory Filtering and Semantic Output Module PTS-SL can be used to obtain the final personalized trajectory from candidate trajectories and make intelligent arbitration for decision-making.

[0200] 1. Using the following trajectory scoring function, obtain a comprehensive score for the K retained candidate trajectories (K≤5, where K represents the number of candidate trajectories):

[0201] The three weighted averages are as follows: w1 represents safety; w2 represents driving style matching; w3 represents comfort; w1 is the safety weight, which measures the importance of avoiding collision risks and obeying traffic rules; w2 is the driving style matching weight, which measures the importance of the match between driving behavior and the user's personalized preferences (such as aggressive, conservative, or conventional) or ride comfort; and w3 is the comfort weight, also known as the smoothness weight, which measures the importance of the kinematic quality of the trajectory itself.

[0202] The above This represents the safety score, the highest priority, and integrates metrics such as the minimum distance (normalized to the 0-1 range) between the trajectory and all static and dynamic obstacles, and the time to collision (TTC), to mitigate collision risks and ensure compliance with traffic rules. The weight w1 is set to a large value to ensure that safety is the most important requirement under any circumstances.

[0203] The above The style matching score is calculated by taking the cosine similarity between the trajectory and the target driving style characteristics (DSE style). It reflects the degree to which the trajectory's acceleration, jerk, lateral sway, and other indicators match the user's personalized preferences (such as aggressive or conservative) or current riding mode. Increasing the weight w2 can make the system output a trajectory that better matches the user's long-term preferences or current settings.

[0204] The above The smoothness score, representing the reciprocal of the variance of the lateral acceleration of the trajectory (a smaller reciprocal indicates smoother performance), is used to evaluate the continuity and smoothness of geometric properties such as curvature and rate of change of curvature of the trajectory. A smooth trajectory reduces unnecessary mechanical wear and provides a more comfortable riding experience. Weight w3 focuses more on the inherent physical properties of the trajectory.

[0205] By adjusting the three weights (w1, w2, w3), the three objectives of safety, personalized driving style, and ride comfort can be balanced.

[0206] 2. Configure at least four output modes as shown in Table 1 below: Table 1 shows the correspondence between patterns and selection strategies.

[0207] 3. Select from K candidate trajectories to obtain the final predicted trajectory, and automatically generate natural language labels based on the following rule engine: If the driving style match score (StyleMatch) is greater than 0.8 and the intention is to change lanes, a natural voice label will be automatically generated indicating that the driver is preparing to change lanes on the left to give you an opportunity to overtake.

[0208] If the driving style match score (StyleMatch) is greater than 0.7 and the speed decrease rate is greater than 0.3, a natural voice label will be automatically generated indicating that the vehicle has slowed down in advance, thus allowing you to maintain a safe following distance.

[0209] Otherwise, the automatic natural language tag will be "Keep in the current lane and drive smoothly".

[0210] 4. The following API interface is used as the output interface and gRPC service is used as the communication protocol, as defined below: / / Trajectory response information{ Repeated trajectory point sequence = 1; / / One point every second for the next 5 seconds .

[0211] The confidence level for a single-precision floating-point number (float) is 2. / / Prediction Confidence .

[0212] The string "driving intent" is set to 3. / / Prepare to change lanes to the left, and slow down urgently, etc. .

[0213] The string "driving style match score = 4" indicates that the driving style matches the given value. / / High match (style trait DSE is radical) .

[0214] The string uses pattern 5; / / Personalization First .

[0215] The index of the selected candidate trajectory is 6 (32-bit signed integer int32). / / Select the index of the candidate trajectory Number}

[0216] Furthermore, the end-to-end latency is ≤ 120ms. This satisfies ISO 26262 ASIL-B (ISO 26262 Automotive Safety Integrity Level B, the automotive safety integrity level B in the ISO 26262 standard for functional safety of road vehicles), thus meeting the corresponding latency requirements.

[0217] Specific implementation examples of the embodiments of this application in practical applications are shown below: The application scenario is illustrated using an example of an emergency lane change by a vehicle at an elevated highway ramp, where the guide lines are obscured by construction barriers. Specifically, the assisted driving vehicle is traveling at 22 m / s (79 km / h) on an elevated highway ramp. A car 30 meters ahead suddenly changes lanes to the left due to ramp congestion, blocking the lane in front of the assisted vehicle. At this time, the guide lines on the high-precision map are temporarily obscured by construction, and the lane detection module outputs noise lines with a false detection rate >40%. The system needs to make a decision within 0.5 seconds to avoid a rear-end collision or emergency braking.

[0218] Step 1: Sample the following content and perform perceptual input and style feature DSE update: The above sensory inputs are as follows: BEV Feature Map: An obstacle such as the vehicle in front is detected 15m directly in front of the vehicle, and there is an empty lane 5m to the left front.

[0219] Guide line detection: Outputs 3 noise lines, all of which deviate from the actual lane center by more than 1.2m.

[0220] Radar data: Vehicle ahead speed 18m / s, relative speed of this vehicle 4m / s, TTC=3.75s.

[0221] The above style features have been updated in DSE as follows: In the 30 seconds prior to the current time, the driver completed 3 lane changes, with the peak average lateral acceleration reaching [value missing]. The rate of change of steering wheel angle is >15° / s.

[0222] The output of the driving style embedding extraction module DSEE is as follows: style feature DSE=[0.92, 0.15, 0.21, ..., 0.88], which is identified as S1, i.e., aggressive type, by the discriminator of the driving style embedding extraction module DSEE, with a confidence level of 94%.

[0223] Step 2: Use the following to generate a VWG-SM aiming point.

[0224] The environmental features E are input into the BEV-Transformer to extract key semantics, namely, there is an obstacle ahead, the left lane is empty, and there is a guardrail on the right. After injecting the style features DSE, VWG-SM generates the 5 sets of aiming points shown in Table 2 below: Table 2 shows five sets of aiming points sampled from the diffusion model.

[0225] Step 3: Using the following methods, the dynamic sector feasible region of the dynamic sector spatiotemporal hard constraint module DFFR is obtained.

[0226] First, given the current vehicle speed v = 22 m / s and the foresee time T_foresee = 3.0 s, we obtain the radius R = 66 m and the angle range θ = 30°. (1+0.01 22)≈36.6°. Half of 36.6° is 18.3°.

[0227] Secondly, based on the offset of the target point from the vehicle center and the heading angle deviation, along with the angle range, it is determined whether to retain the predicted trajectory: If the target point of the predicted trajectory W1 is offset from the center of the vehicle by 60m and the heading angle deviation is 3.47° < 18.3°, then the predicted trajectory W1 is determined to be retained.

[0228] If the target point of the predicted trajectory W2 is offset from the center of the vehicle by 65m and the heading angle deviation is 0.43° < 18.3°, then the predicted trajectory W2 is determined to be retained.

[0229] If the target point of the predicted trajectory W3 is 61m off from the center of the vehicle and the heading angle deviation is 3.29° < 18.3°, then the predicted trajectory W3 is determined to be retained.

[0230] If the target point of the predicted trajectory W4 is offset from the center of the vehicle by 13m and the heading angle deviation is 21° > 18.3°, then the predicted trajectory W4 is determined to be removed.

[0231] If the predicted trajectory W5 is reversed, then the predicted trajectory W5 is determined to be removed.

[0232] Ultimately, three predicted trajectories, W1, W2, and W3, were retained as candidate trajectories.

[0233] Step 4: Use the following to perform personalized trajectory selection for the personalized trajectory filtering and semantic output module PTS-SL.

[0234] With a safety weight of W1=0.3, a driving style matching weight of W2=0.35, and a comfort weight of W3=0.35, the three remaining candidate trajectories are comprehensively scored to obtain the scores of each candidate trajectory, as shown in Table 3 below.

[0235] Table 3 Scores for each candidate trajectory

[0236] Based on the trajectory with the highest overall score, W1 is selected as the final predicted trajectory. At this point, W1, with its target being a left lane change, is highly consistent with the aggressive style characteristic DSE.

[0237] Step 5: Use the following content to perform semantic output and execution.

[0238] The output instructions are as follows: { "Personalized trajectory trajectory": [[15.2, 1.8], [18.5, 2.1], [21.8, 2.4], [25.1, 2.7], [28.4, 3.0]], "Confidence" score: 0.81 "Driving Intention": "Preparing to change lanes to the left to give you an overtaking opportunity." "Driving style alignment": "High alignment (DSE is aggressive)" "Usage mode_used": "Personalization preferred"}.

[0239] Then, the personalized trajectory is input into the model prediction controller to generate steering angle and acceleration commands. The vehicle smoothly changes lanes to the left and completes the overtaking maneuver without sudden braking or shaking.

[0240] Step 6: Use the following methods to verify the effect and provide closed-loop feedback.

[0241] This application embodiment verifies the performance offline. While the vehicle executes the final predicted trajectory, at least one of the driving status and driving experience during the execution of the final predicted trajectory is collected in real time to obtain the actual usage effect. The actual usage effect is then compared with the baseline model to verify the effect and generate the verification result for this driving test.

[0242] The actual usage effect is compared with the baseline model based on the following verification metrics to generate the verification results of this test. These verification metrics may include, but are not limited to, at least one of the following: baseline, model convergence speed, user rating, and model convergence speed.

[0243] For example, due to the noise of the guide line, the output trajectory decelerates in place, causing the following vehicle to brake suddenly, increasing the probability of emergency braking by 47%.

[0244] In this embodiment of the invention, the lane change was successful without emergency braking, and the passenger comfort rating (subjective questionnaire) changed from 3.2 / 5 to 4.8 / 5.

[0245] The model convergence speed is as follows: with the same amount of data, this system achieves 78% minADE (minimum average displacement error) after 100 training rounds, while the baseline is only 63%.

[0246] Based on the above verification results, the following system iterations will be performed: First, since the driver's subsequent three lane changes all occurred in the same scenario, the system automatically increased the DSE (Driving Style Feature) weight of the left lane change strategy by 5%. This is achieved by continuously analyzing the driver's actual driving behavior, identifying repetitive actions that do not match the current default strategy, and judging real-time changes in driver habits. When trigger conditions are met, the system adjusts the weight parameters reflecting the driver's habits accordingly, thereby updating and generating a personalized trajectory.

[0247] Secondly, a new version of the trajectory scoring function VWG-SM v1.1 was generated. After retraining the trajectory generation model, the lane-changing success rate was improved to 91%.

[0248] In the embodiments of this application, driving style embedding, strategy activation, physical constraints, diversity optimization and personalized selection are integrated to form a complete, trainable and deployable end-to-end system.

[0249] This application provides a vehicle including one or more processors for implementing the predicted trajectory generation method described above. The vehicle may include, but is not limited to, one or more of heavy-duty trucks, commercial vehicles, passenger vehicles, and special-purpose vehicles.

[0250] The trajectory prediction generation method of this application embodiment is applied to an electronic device. The aforementioned electronic device can be a PC (Personal Computer) device. The PC device can include, but is not limited to, desktop computers, tablets, or laptops. An input information interface is set on the webpage or APP (application) page of the PC device. Thus, by displaying the input information interface, users can fill in information to obtain the user's required information and achieve the user's purpose.

[0251] Of course, this electronic device may also include, but is not limited to, an in-vehicle terminal connected to the vehicle. The in-vehicle terminal connected to the vehicle may be, but is not limited to, a vehicle controller, body processor, center console, or automotive HUD (Head-Up Display).

[0252] Figure 4 The diagram shown is a structural schematic of the electronic device 50 provided in an embodiment of this application.

[0253] like Figure 4 As shown, the electronic device 50 includes one or more processors 51 for implementing the predicted trajectory generation method as described above.

[0254] In some embodiments, electronic device 50 may include storage medium 59. For example, computer-readable storage medium may store a program that can be invoked by processor 51, and may include non-volatile storage medium. In some embodiments, electronic device 50 may include memory 58 and interface 57. In some embodiments, electronic device 50 may also include other hardware depending on the specific application.

[0255] The computer-readable storage medium of this application embodiment stores a program that, when executed by processor 51, is used to implement the predicted trajectory generation method described above.

[0256] This application provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the method described in any of the preceding claims.

[0257] This application also provides a computer program stored in a computer-readable storage medium, for example... Figure 4 The storage medium 59, and when the processor executes the computer program, causes the processor 51 to perform the method described above.

[0258] This application may take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing program code. Computer-readable storage media include permanent and non-permanent, removable and non-removable media, and information storage can be implemented using any method or technology. Information may be computer-readable instructions, data structures, program modules, or other data. Examples of computer-readable storage media include, but are not limited to: phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0259] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Without further limitation, an element qualified by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A method for predicting trajectories, characterized in that, include: Based on the vehicle's environmental perception information, predict the noise level of the environmental perception information; Based on the environmental perception information, the noise is removed, and multiple sets of anchor points are generated; The multiple sets of aiming points are used to represent multiple predicted trajectories in multiple directions; The final predicted trajectory is determined from multiple predicted trajectories.

2. The method for generating predicted trajectories as described in claim 1, characterized in that, Determining the final predicted trajectory from multiple predicted trajectories includes: Based on the real-time vehicle driving status in the environmental perception information, a dynamic sector feasible region for constraining candidate points is determined; Based on the constraint of the dynamic sector feasible region as the trajectory generation, the trajectory that does not exceed the dynamic sector feasible region among the multiple predicted trajectories is taken as the candidate trajectory containing the candidate point. The final predicted trajectory is determined from the candidate trajectories.

3. The method for generating predicted trajectories as described in claim 2, characterized in that, The step of predicting the noise of the environmental perception information based on the vehicle's environmental perception information includes: Based on the external environment perception information and real-time vehicle driving status in the environmental perception information, determine the environmental characteristics and predict the noise of the environmental perception information. Based on the environmental features, the noise is gradually removed by the aiming point decoder to obtain multiple sets of initial aiming point sequences; The multiple sets of initial aiming point sequences are post-processed and smoothed to obtain multiple sets of aiming points that meet the requirements of continuous motion and have clear intention semantics.

4. The method for generating predicted trajectories as described in claim 2, characterized in that, The step of predicting the noise of the environmental perception information based on the vehicle's environmental perception information includes: Based on the external environment perception information and real-time vehicle driving status in the environmental perception information, determine the user's driving style characteristics, environmental characteristics, and predict the noise of the environmental perception information. Based on the environmental perception information, the noise is removed, and multiple sets of anchor points are generated, including: Based on the user's driving style characteristics and environmental characteristics, the noise is removed, and multiple aiming points are generated. Determining the final predicted trajectory from the candidate trajectories includes: Each candidate trajectory is scored based on a comprehensive evaluation of safety, driving style compatibility, and comfort. Based on the scores of each candidate trajectory, the final personalized trajectory is determined from the candidate trajectories.

5. The method for generating predicted trajectories as described in claim 4, characterized in that, Based on the user's driving style characteristics and environmental characteristics, the noise is removed, and multiple sets of aiming points are generated, including: Based on the environmental and driving style features, the noise is gradually removed by the aiming point decoder to obtain multiple sets of initial aiming point sequences; The multiple sets of initial aiming point sequences are post-processed and smoothed to obtain multiple sets of aiming points that meet the requirements of continuous motion and have clear intention semantics.

6. The method for generating predicted trajectories as described in claim 3 or 5, characterized in that, The method further includes: the aiming decoder contains a trained noise predictor, which predicts the noise of the environmental perception information based on the vehicle's environmental perception information; the trained noise predictor is obtained in the following manner: Obtain the true ideal anchor point sequence in the current vehicle scene, as well as the environmental and style features of the current vehicle scene; According to each time step, real noise is added to the ideal anchor point sequence to determine the noisy anchor point sequence for each time step; The noisy aiming point sequence, the environmental features, and the style features are input into the noise predictor to be trained, so as to output predicted noise. Using backpropagation, the model parameters of the noise predictor to be trained are updated, and the training iteration is repeated until the error between the predicted noise and the real noise decreases to the end of the iteration, thus obtaining the trained noise predictor.

7. The method for generating predicted trajectories as described in claim 4, characterized in that, The candidate trajectories are then comprehensively scored based on safety, driving style compatibility, and comfort to obtain a score for each candidate trajectory, including: Each candidate trajectory is weighted according to its safety score and safety weight, trajectory smoothness and comfort weight, and the degree of matching of driving style and driving style matching weight to obtain a score for each candidate trajectory. The step of determining the final personalized trajectory from the candidate trajectories based on their scores includes: From the scores of each candidate trajectory, the candidate trajectory with the highest score is selected as the personalized trajectory.

8. The method for generating predicted trajectories as described in claim 4, characterized in that, The step of determining the user's driving style characteristics based on the external environment perception information and real-time vehicle driving status in the environmental perception information includes: The external environment perception information and real-time vehicle driving status are input into the driving style embedding and extraction module to output the user's driving style features; the driving style embedding and extraction module is trained using a style label dataset.

9. The method for generating predicted trajectories as described in claim 8, characterized in that, The driving style embedding and extraction module includes: a preprocessing module, an intermediate layer, and a time attention mechanism; The preprocessing module performs up-dimensional mapping based on the external environment perception information and real-time vehicle driving status in the environmental perception information to obtain the features of each time step. The intermediate layer fuses the context information of adjacent time steps based on the features output by the preprocessing module and extracts the hidden state of each time step of the fused context information. The weights of each time step are learned through a time attention mechanism, and the hidden states of each time step are weighted and summed with the corresponding learned weights of each time step to generate a third-style embedding vector.

10. The method for generating predicted trajectories as described in claim 2, characterized in that, The step of determining the dynamic sector-shaped feasible region for constraining candidate points based on the real-time vehicle driving status in the environmental perception information includes: Based on the vehicle's current driving status and its own dynamic parameters, a dynamic sector-shaped feasible area is delineated. Under the constraint of using the dynamic sector feasible region as the trajectory generation method, the trajectory among the multiple predicted trajectories that does not exceed the dynamic sector feasible region is selected as the candidate trajectory containing the candidate point, including: Based on the constraints of the dynamic sector feasible region as the trajectory generation, all trajectory points of the predicted trajectories that exceed the maximum lateral acceleration and allowable curvature range of the vehicle are deleted, and the trajectories that do not exceed the dynamic sector feasible region in the predicted trajectories are obtained as candidate trajectories. The boundary of the dynamic sector feasible region is determined by the maximum lateral acceleration and allowable curvature corresponding to the current vehicle speed.

11. The method for generating predicted trajectories as described in any one of claims 2 to 5 and 7 to 10, characterized in that, The step of determining the dynamic sector-shaped feasible region for constraining candidate points based on the real-time vehicle driving status in the environmental perception information includes: The current coordinates of the vehicle's rear axle are used as the origin of the sector. Based on the instantaneous speed of the vehicle and the configurable look-ahead time, the dynamic radius of the sector and the total angle of the sector are obtained; Based on the origin of the sector, the dynamic radius of the sector, and the total angle of the sector, an adjustment is made to obtain a dynamic sector feasible region for constraining candidate points.

12. The method for generating predicted trajectories as described in claim 11, characterized in that, The method further includes: When a collision risk is detected in the vehicle, the look-ahead time is shortened, the dynamic radius of the sector is reduced, and the total angle of the sector is narrowed. The adjustment based on the origin of the sector, the dynamic radius of the sector, and the total angle of the sector to obtain a dynamic feasible region of the sector for constraining candidate points includes: The feasible region of the dynamic sector is generated based on the origin of the sector, the total angle of the narrowed sector, and the reduced dynamic radius of the sector.

13. The method for generating predicted trajectories as described in any one of claims 2 to 5 and 7 to 10, characterized in that, Under the constraint of using the dynamic sector feasible region as the trajectory generation method, the trajectory among the multiple predicted trajectories that does not exceed the dynamic sector feasible region is selected as the candidate trajectory containing the candidate point, including: For each candidate point of a multiple candidate trajectory, determine whether the distance from the current position is not greater than the dynamic sector feasible area, and whether the vehicle's self-heading angle deviation is not greater than half of the angle range at the current moment. If the conditions are not met, the unqualified trajectory is discarded; If the conditions are met, the qualified trajectory is retained as a candidate trajectory.

14. The method for generating predicted trajectories as described in any one of claims 2 to 5 and 7 to 10, characterized in that, The method is applied to a trajectory generation model; the method further includes: during the training of the trajectory generation model, adding a corresponding penalty term to the candidate trajectories that are close to the boundary; the penalty term is used to learn to avoid the candidate trajectories that are close to the boundary during the training of the trajectory generation model, and the penalty term is negatively correlated with the distance between the candidate trajectory and the boundary of the dynamic sector feasible region; And / or, The trajectory generation model includes a trained trajectory distribution penalizer, and the method further includes: using the trained trajectory distribution penalizer to remove noise based on the environmental perception information and generate multiple sets of anchor points; the trained trajectory distribution penalizer is obtained in the following manner: Based on the environmental characteristics and driving style characteristics of the vehicle in the current environment, the aiming point sequence with added noise is de-noised in reverse to obtain multiple sets of initial aiming point sequences; The multiple sets of initial anchor point sequences are input into the trajectory distribution penalizer to be trained, so as to output the distribution likelihood corresponding to each set of initial aiming point sequences; Based on the distribution likelihood, a diversity loss term is determined; the diversity loss term is used to increase the diversity of the multiple predicted trajectories under the same input conditions; the diversity loss term is negatively correlated with multiple sets of initial aiming point sequences whose differences are less than a preset difference; Using backpropagation, the model parameters of the trajectory distribution penalty device to be trained are updated, and the training iteration is repeated until the loss of the trajectory distribution penalty device to be trained meets the iteration termination condition, thus obtaining the trained trajectory distribution penalty device.

15. An electronic device, characterized in that, It includes one or more processors for implementing the predictive trajectory generation method as described in any one of claims 1 to 14.

16. A vehicle, characterized in that, It includes one or more processors for implementing the predictive trajectory generation method as described in any one of claims 1 to 14.

17. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, implements the predictive trajectory generation method as described in any one of claims 1 to 14.

Images